Nonlinear solver misses one test profile (test data seed 0) #134

With the acceptance test data generated by create_test_data(seed=0) (seed parameter from #131), the nonlinear solver reconstructs image 3 (linear profile 0.15·z plus noise) with rmse 0.178 against the input. The linear solver on the same data gives 0.014.

For seeds 1–7 the nonlinear rmse is at most 0.036, so this is a single reproducible weak case, not a general accuracy problem. The suite now uses seed 1; seed 0 can be used to investigate. Possible causes I have not checked: a convergence or bounds issue of the TNC minimisation, or the weighting terms (weighting_preference = -6e-4, weighting_curvature = 1e-7, num_iterations = 2000 in the test config).

Reproduce: create_test_data(seed=0), then S1, S2, -coord, -s3_fast, and --analysis with the nonlinear solver.